Uncertainty in Robotics
We are Roboteers Club GNI from Hyderabad. Our goal is to spread awareness on robotics and its applications.
Uncertainty is a major determinant of success in robotics applications. Online policy search algorithms need to balance exploration and exploitation to reach maximum learning speed. In robot vision, algorithms need to be robust against occlusions, blur, and light reflections. The inherent stochasticity and scarcity of human demonstration data needs to be handled in skill transfer. Many of the machine learning tools in use with robotics applications deliver highly accurate predictions when provided sufficient data, however their reliability in the face of uncertainty remains substantially below the requirements of safety-critical application areas such as healthcare robotics and autonomous mobility. Such applications demand machine learning components that are robust against missing or noisy information, and can accompany their predictions with calibrated confidence scores that can be reliably used in downstream operations.
Modern robotic systems use machine learning tools in various stages of their execution spanning from perception to control. While these tools enhance the capabilities of the robotic systems, they also increase their sensitivity to situations that are underrepresented in the training data. Recent advances in machine learning provide many new methodologies to mitigate such sensitivities. The most critical performance bottlenecks in machine learning applications to robotics are yet to be identified and principled approaches that mitigate them remain to be developed. The goal of this research topic is to bring together the expertise of machine learning researchers in finding algorithmic solutions to generic data science problems and the domain expertise of robotics researchers to overcome the barriers that hinder reliable deployment of cutting-edge machine learning technologies into safety-critical robotics use cases. The contributions of this special issue will serve as a step towards enabling new robotic technologies in new areas of our lives that are excluded thus far due to safety concerns.
This Research Topic welcomes original ideas that employ the capabilities of probabilistic methods for solving key problems of robotics in unique and thought-provoking ways. We highly value simple and elegant algorithmic advances provided that their success is empirically demonstrated in sufficient detail. We also greatly encourage theoretical contributions reporting analytical results obtained by formal methods that unravel unknown properties of existing models. We prioritize contributions in the topics below, although articles in other related areas will also receive full consideration:
• Uncertainty-aware robot vision,
• Safe exploration in policy search,
• Manipulation with partial observations,
• Human-to-robot skill transfer,
• Adaptive optimal control,
• Sample-efficient reinforcement learning,
• Dynamical systems modelling,
• Data scarcity in robotics.
